how I build stunning websites with AI

Hey folks, it's Monday. Sorry about that.

We had some really incredible releases over the last week, so I'm glad that I got to squeeze them into this newsletter.

It was also my birthday on the 19th of September, and I turned a whopping 22. I can already feel the arthritis creeping in.

📌 TL;DR

• One Claude → Chat and Cowork finally live in one conversation.

• Jev → A new really cool AI model

• Higgsfield’s API → Pay only for what you generate, not tokens you may waste.

• Notion Skills → Notion is becoming Switzerland for AI.

• Claude Code Projects → One conversation coordinates cloud agents after you close your laptop.

• Builder’s notes → my workflow for vibecoding beautiful landing pages and 104K-view carousels in Paper

Cowork and chat become one Claude

Anthropic is finally merging regular chat and Cowork into one Claude.

This was such an obvious design decision that it should have been there from day one. Nobody knew which box a task belonged in, and Claude should have been smart enough to work that out itself.

The simplest distinction is that Claude Chat is for conversations. Cowork and Claude Code are agents, meaning they can choose and take actions for you.

Cowork is the friendlier version. Claude Code still gives you more power and flexibility, which is why I prefer it, but Cowork probably covers 90% of what most people need.

Jev is really cool: a new kind of AI model that makes decisions instead of writing

The guy who invented a kind of well-known product called ChatGPT, has spent the last 2 years in stealth building a new way to train models.

He’s released it to the world and it’s called Jev.

The internet is losing its mind over it. here is the hype explained in 30 seconds...

The LLMs we know and love basically write an essay word by word. That flexibility is great for complex jobs, but slow and expensive when software needs thousands of small decisions.

Jev does not write. Give it a situation and a fixed list of possible answers, and it picks one with a confidence score.

Give it an email, for example, with the options “sales lead, support, newsletter or spam”.

It chooses a category and tells you how confident it is.

There are three reasons people are so excited.

1. It technically has a 0% hallucination rate as it cannot make up answers

2. it is disgustingly faster and cheaper (70–500ms per decision, with input costing just 4.2¢ per million tokens and output free.)

3. Computer use with Jev is nuts

A browser agent usually asks a language model to study the page, think through the task and write its next action. Then it repeats that process after every click.

With Jev, the browser turns what is currently on the page into structured information: the available buttons, fields and actions.

Jev rapidly chooses the next action and target from that allowed list, the browser carries it out, then the loop starts again on the updated page.

Jev does not need to see the page as pixels in this setup. It receives a structured map of the controls.

You can see more of the model and computer-use examples in Matthew Berman’s breakdown.

Greg Isenberg’s interview gets into practical systems like inbox classification and lead scoring.

And then my brain goes to all the out-of-the-box possibilities this unlocks.

I think this is a really, really cool breakthrough. Computer use is probably the clearest practical example right now, but this could end up everywhere.

Higgsfield just killed Higgsfield?

Higgsfield just launched a pay-as-you-go option, and this makes so much more sense for anyone who does not create content every day.

A normal Higgsfield subscription charges you monthly for a fixed allowance, with a pretty large markup on the tokens underneath.

The problem is that AI content usage is never consistent. One month you might burn through the whole allowance in a week. The next, you barely touch it.

With the new API, you top up a balance and only pay for the images and videos you actually create.

I absolutely love this release but the launch creative might be more interesting than the API itself.

Higgsfield says this launch video was made entirely with AI.

I could not tell at all.

This is the state of AI video now.

Notion is becoming the Switzerland of AI

Someone said that Notion is the Switzerland of AI and I can’t stop thinking about it.

The AI models are so smart now, the biggest bottleneck is context, meaning what they know about you.

During World War II, Switzerland managed to remain neutral ground because people from across a divided Europe kept their money in Swiss banks.

And right now… your AI needs memory, and it needs somewhere it can read from and write to, so people just connect Notion.

As a bonus, the models seem very good at using it.

This week, Notion made that strategy even clearer with Notion Skills.

Skills are reusable instructions for your agents. think SOP for AI agents.

You store and update them in Notion, then use them with Claude, ChatGPT, Codex or whatever agent comes next.

So Notion positions itself as the database, which it always was, and emphasises that it is the permanent part of your AI setup.

I have already started moving my own skills into Notion. It may make the separate Skills Atlas I built for sharing them with my team less necessary, and I am completely okay with that.

Grokbot, Claude, Codex and Muse can all change around it while they interact with the same faithful and reliable base: Notion.

Good strategy.

a viable solution to claude code brainrot

Anthropic is giving Claude Code a new Projects feature, and I am so, so excited for this way of working.

I have not used it yet because it still has not appeared in my desktop app, but the idea is simple: one Project becomes the permanent home and coordinator for a larger piece of work.

Say I am building a new website. Right now I might have five different sessions open:

• one for the copywriting

• one for the branding

• one planning the sections

• one building the site

• one working out the checkout

Then I spend the day bouncing between them, trying to remember what each one knows and passing information around manually.

It is basically TikTok but for work. It rots my brain.

With Projects, I would talk inside one continuous website conversation. That Project holds the shared goal, context and files, then coordinates the separate work behind the scenes.

Each thread it starts is a full Claude Code cloud session doing one part of the job.

So the Project might send the copy to one thread and the build to another, keep track of both, then show me what has finished or needs an answer. The threads keep running in the cloud after I close my laptop.

I cannot wait for this to reach my account. Managing ten different AI tasks has somehow become a job of its own, and this feels like a much better way to work.

Also this week...

• Google launched Gemini 3.8 Live and Extended Thinking → Live is the fast, lower-cost model for conversations at scale, while Extended Thinking handles harder jobs that need several reasoning steps.

• ElevenLabs launched Reception, an off-the-shelf AI receptionist → Give it a dedicated number or forward your existing line, and it answers calls, handles common questions, checks availability and books appointments. Afterwards, you get the summary, transcript and recording.

• Salesforce in Claude is now in beta → The plugin gives sales teams 37 ready-made skills that pull CRM records and Slack deal context into Claude. It can prep meetings, review deals, schedule daily briefs, build plans and draft CRM updates for approval.

• Claude Code now reads AGENTS.md → FINALLYYYY

• Grok Bot adds voice → You can now have a spoken back-and-forth with the agent instead of typing every instruction.

💡 Builder's notes

How I vibecoded a really nice landing page for The AI Course

I built the entire The AI Course landing page with Claude Code and Codex inside Paper.

I planned every section, worked out the job it needed to do, then wrote most of the copy.

For each section, I asked Claude or Codex for around 3-5 completely different design directions.

I reviewed them, explained exactly what worked and didn't, then had it turn that feedback into another three variations.

I kept repeating that loop until the section felt right, then I would move on to the next.

and then finally, once I had all those sections done, I told Claude to stitch them all together and publish it live

in the process, I found GPT Astra to be at least on par with Fable for design work so far, sometimes better.

I've also been making Instagram carousels in Paper.

I've also been using Paper for Instagram carousels.

One of those carousels hit 104K views.

If you're building websites, landing pages or carousels with an agent, Paper is a must-have.

I used an illustration to understand what my agents built.

It's easy to build lots of things with AI while having no idea how they work (which is an issue).

That happened with a new client-onboarding flow I built.

I used an Annotated Illustrations skill to turn the whole system into a hand-drawn explanation.

It mapped the order things happen in, how the parts relate, the conditions that change what happens next, and anything it couldn't verify.

and here were some others…

Instead of reading through a pile of code and documentation, I had one picture that showed me what my agents had actually built.

🧰 Tools to try

No tools this week. I didn't find anything worth your time.

🥣 Brain food

• Applied AI Doesn’t Work → one of the most insightful articles I've read recently about AI implementation

• The infinite FAQ → Such a neat little website interaction, I will 100% be implementing this on all future websites

• Replacing the browser with ChatGPT desktop → A fascinating demo of how some users are containing all of their work within the ChatGPT desktop app

• Meta Quest sends a 3D model to a physical printer → It's actually fucking crazy to think this is a real video. The future is actually here.

• Premium free-font roundup → inspiration for your next design project